МОДЕЛІ ГЛИБИННОГО НАВЧАННЯ ДЛЯ ПРОГНОЗУВАННЯ РУХУ ЛЮДИНИ У ВІДЕОПОТОКАХ

Context. The problem of accurately predicting human movement in an environment is critical for applications in monitoring, search, and navigation systems. Existing approaches often struggle to integrate spatial and temporal dynamics of trajectories while processing real-time video streams.Objective. The goal of this work is to develop a deep learning-based framework capable of predicting human motion by combining object-level features and spatio-temporal trajectory information extracted from video streams.Method. The proposed method integrates YOLO11 for object detection, which extracts coordinates, velocity, movement direction, and position relative to the environment. A graph neural network models local and global relationships between environment nodes, aggregating features while considering terrain structure and obstacles. Spatio-temporal attention highlights the most relevant moments in the trajectory, enhancing prediction accuracy. The model processes sequences of frames from video streams to predict subsequent positions of each tracked object in real time.Results. Experiments on video sequences with varying motion scenarios, trajectory lengths, and speed variations demonstrated high prediction accuracy. The proposed method effectively integrates spatial and temporal features, outperforming baseline models in tracking and motion prediction tasks.Conclusions. The results confirm that the proposed deep learning framework is suitable for real-time human motion prediction in complex environments. Future research may focus on extending the approach to multi-agent scenarios, optimizing computational performance, and testing on larger and more diverse datasets